Analysis of a German Legal Citation Network
Tobias Milz
a
, Michael Granitzer
b
and Jelena Mitrovi
´
c
c
Chair of Data Science, University of Passau, Innstraße 41, 94032 Passau, Germany
Keywords:
Legal Language Processing, NLP, Network Analysis, PageRank, Neo4j, TF-IDF, Node Similarity.
Abstract:
The paper introduces the creation and analysis of a German legal citation network. The network consists of
over 200.000 German court cases from all levels of appeal and jurisdiction and more than 50.000 laws. Ref-
erences to court decisions and laws are extracted from within the decision text of the court cases and added
as links to the network. We apply network-based analysis techniques to support common legal information
retrieval tasks such as identification of important court decisions and laws and case similarity searches. Fur-
thermore, we demonstrate that the German case citation network displays scale-free behaviour, similar to that
of the U.S. and Austrian Supreme Courts as shown by previous research.
1 INTRODUCTION
Investments into supporting technologies for the le-
gal industry (also known as Legal Tech) have set a
new record high in 2019. The market value was es-
timated to be at 17.32 billion U.S. dollars and is ex-
pected to reach 25 billion by 2025
1
. This growing
interest is also apparent in the legal research com-
munity, which, especially in recent years, has given
a surge of attention to the use of AI in the various
areas of technology-assisted legal research. One of
these areas is concerned with the development of NLP
based information retrieval systems, which has been
of particular importance to this domain since the early
1960s (Widdison, 2002). This can be attributed to the
ever growing amount of text-based information the
law profession produces.
Research in the legal realm reveals challenges as it
not only differs based on the underlying legal system
(civil law, common law, customary law, etc.) but also
between the countries with the same or similar legal
foundation. NLP based systems have the task to ad-
just accordingly and investigate which studies can be
translated and replicated across these different mani-
festations of the legal domain.
In the early nineteenth century, the use of legal
citation indexes started to become an important infor-
a
https://orcid.org/0000-0003-3159-7666
b
https://orcid.org/0000-0003-3566-5507
c
https://orcid.org/0000-0003-3220-8749
1
https://www.statista.com/statistics/1155852/legal-
tech-market-revenue-worldwide/
mation retrieval tool in case law (Geist, 2009). Know-
ing which precedents are still being cited, or which
new cases are gaining a lot of attention was crucial
information - information that was otherwise diffi-
cult to obtain due to the increasing amount of new
cases. This practice could be seen as an early exam-
ple of citation analysis. Today, the study of citation
networks is an established research area across many
domains including citation behaviour analysis of sci-
entific publications (Milz and Seifert, 2018). Conse-
quently, legal scholars have long been theorising the
benefits of using citation networks and network anal-
ysis algorithms (Neale, 2013) for legal research pur-
poses. For example, case citation networks can reveal
critical information on precedents (Cross et al., 2010)
based on the characteristics of the citations towards
the precedent candidates. Although this research has
received some consideration from legal scholars in the
past, as section 2 will outline, we believe it is still un-
derutilised, especially for civil law systems and the
Germanic law in particular.
In this publication, we describe the development
and analysis of a German legal citation network. In
contrast to most previous research, however, we also
include citations to laws as opposed to common case-
to-case citation networks. We analyse the results from
a quantitative standpoint and ignore in-depth legal in-
terpretations as this is out of the scope of this study.
However, by presenting our approach and findings,
we hope to create a baseline for future analysis and
stir the interest of legal scholars to interpret the re-
sults.
Milz, T., Granitzer, M. and Mitrovi
´
c, J.
Analysis of a German Legal Citation Network.
DOI: 10.5220/0010650800003064
In Proceedings of the 13th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2021) - Volume 1: KDIR, pages 147-154
ISBN: 978-989-758-533-3; ISSN: 2184-3228
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
147
Our paper is structured as follows: In section 2 we
summarise some of the related work in this field and
indicate how our work can provide new insights into
this area of research. In section 3 we introduce the
data collection and citation extraction process before
presenting the details of the resulting citation network
in section 4. In section 5 and 6 we present the results
of the network analysis and conclude our overall con-
tribution.
2 RELATED WORK
As indicated in section 1 the use of case citation net-
works for legal information retrieval has a long his-
tory. Some of the most important studies in this area
are mentioned here in chronological order.
One of the first and most extensive efforts on le-
gal network analysis research to date was conducted
by Thomas A. Smith in 2005 (Smith, 2005). In this
study, he acquired a citation network containing over
4 million U.S. federal and state cases and compared
its properties with the world wide web and a physics
literature citation network. His work provides a com-
prehensive overview of use cases for legal citation
network analysis. In 2007 (Fowler et al., 2007) anal-
ysed a network containing 26,681 opinions from the
U.S. Supreme Court. They presented a method for
identifying current and future precedent cases using
a ranking system based on citation analysis. In 2009
(Geist, 2009) analysed 80,195 opinions from the Aus-
trian Supreme Court of Justice. In his thesis, he pro-
vides evidence for the benefits of using citation counts
as a metric for case relevance rankings.
Another analysis of the citation practices of the
U.S Supreme Court was undertaken by (Cross et al.,
2010). However, in this study, the researchers used
the network to also investigate the citation behaviour
of the individual Justices. In 2011 (Staffan Malm-
gren, 2011) analysed a citation network containing
14,327 decisions from the European Court of Justice
and the General Court. He compares the precision
of multiple network algorithms such as PageRank or
node degree values for identifying relevant cases to
a specific topic. An analysis of cross-country cita-
tion behaviour was conducted by (Gelter and Siems,
2012). They analysed the citations from decisions of
the highest courts from ten different European coun-
tries. In 2013 (Neale, 2013) created a citation network
by extracting citations from 594,540 Canadian court
opinions. Among other findings, this study conducted
a statistical analysis of citations over time and identi-
fied that most cases will not get cited anymore after 3
to 15 years with the exception of decisions from the
Supreme Court of Canada, which still get cited after
50 years.
To the best of our knowledge, one of the first
larger studies on German case citation networks was
conducted by (Corinna Coupette, 2019). However,
the author only considered cases and citations from
and to the federal constitutional court (”Bundesver-
fassungsgericht”). In 2020 (Katz et al., 2020) com-
pared yearly snapshots of law citation networks of
the U.S. and Germany from 1994 to 2018. In con-
trast to previous research, the authors concentrated on
the analysis of laws and the references (citations) be-
tween them. One aggravating problem for the anal-
ysis of German legal data is the lack of freely avail-
able datasets. A problem that has been recognised by
(Ostendorff et al., 2020) who in response made an ef-
fort to collect data from various sources (see section
3.1) and created the largest openly available German
legal dataset to date. In their work, they also intro-
duced a prototype of a citation network and suggested
future work to investigate the network’s properties.
Similarly, other recent datasets have become available
(Urchs et al., 2020; Urchs et al., 2021) to provide an-
notated legal datasets for legal argumentation identifi-
cation. The most recent analysis of German legal cita-
tion networks was conducted by (R
¨
onneburg, 2021).
He created a small citation network based on about
55.000 court cases. One aspect of this study is the
thematic similarity comparison between cases within
automatically detected clusters. Among the two in-
vestigated clusters, the legal expert could confirm a
large semantic similarity.
3 PROBLEM STATEMENT
As highlighted by the size of the datasets in section
2, the largest legal research studies have been con-
ducted on the law systems of the U.S. and Canada.
Research on specific European jurisdictions is still to
receive a similar amount of attention. Furthermore,
previous research on this topic has focused primarily
on case-to-case citation networks from specific courts
(e.g. US Supreme Court, European Court of Human
Rights or Austrian Supreme Court) and case-to-law
citations have not been considered at all. However,
the inclusion of laws in the citation network provides
a more detailed representation of the court’s deci-
sion process. This can help to provide more robust
network-based similarity measures (see section 5.3.1)
and furthermore, allow us to identify the most impor-
tant and influential laws (see section 5.2.2). Conse-
quently, we saw the opportunity and need for a large
scale, descriptive analysis of a German legal citation
KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval
148
network that considers courts and cases from all levels
of appeal. The goal of this study is to analyse inter-
esting properties of the network (e.g. case relevance,
court relations and law importance) from a quantita-
tive point of view (i.e. excluding legal knowledge).
3.1 Data Collection
One reason for the discrepancy between the amount
of North American and European empirical legal re-
search is the lack of openly available datasets of Eu-
ropean court cases. In contrast, ”Court Listener”
2
is
a freely accessible dataset for U.S legal data that con-
tains more than 3.7 million precedential court deci-
sions. In the case of Germany, the only databases that
reach that level of complexity are commercially op-
erated (e.g. Juris GMBH
3
, C.H. Beck Journal
4
). Al-
though, the Federal Ministry of Justice and Consumer
Protection
5
has published about 55,000 court cases
from German federal courts and most states publish
some records on their individual websites, not all of
them allow users to scrape their content. As a result,
until recently, there was no centralised openly avail-
able source for legal data from different jurisdictions.
This fact was recognised by (Ostendorff et al.,
2020) who published an openly available dataset
called ”Open Legal Data”
6
containing over 200,000
court decisions and 50,000 laws from over 1,000 dif-
ferent courts which they collected directly from the
courts and the states and federal websites. However,
as some of the data had been scraped from the indi-
vidual state websites, newly published court decisions
are not always added automatically. Other free le-
gal data services such as Rewis
7
and openJur e.V.
8
on the other hand have either fewer cases or do not
allow access to their data. Consequently, Open Le-
gal Data represents the largest and most accessible le-
gal dataset available at the moment and has therefore
been taken as the foundation for our citation network.
3.2 Citation Extraction
In order to create the links between the nodes in the
citation network, it is paramount to be able to effec-
tively extract the citations within the decision text of
the cases. Unfortunately, this task is not trivial as ref-
erences to previous court decisions or laws are not
2
https://www.courtlistener.com/; 26.04.202.
3
https://www.juris.de/; 26.04.21
4
https://beck-online.beck.de/; 26.04.21
5
https://www.rechtsprechung-im-internet.de; 26.04.21
6
https://de.openlegaldata.io/; 24.04.21
7
https://rewis.io/; 26.04.21
8
https://openjur.de/; 26.04.21
always distinctly identifiable. Additionally, there is
only one annotated dataset for German legal cases
available at the moment (Leitner et al., 2020) and it
is unsuitable for training or testing citation extraction
methods. We will not go into too much detail about
the intricacies of German law referencing but want to
highlight some difficulties that we encountered.
The leading factor for the complications in iden-
tifying case-to-case citations is the lack of a well-
structured unique identifier for court decisions. The
introduction of the European Case Law Identifier
(ECLI) in 2011 might solve this problem in the fu-
ture, but older and current cases are still being refer-
enced via a file number. Unfortunately, the file num-
bers are not as well structured as the ECLI and lack
uniformity, which aggravates the use of automatic ex-
traction methods.
In regards to the law-reference extraction, we sim-
plified this task by only considering citations that are
indicated by the article sign (”§”). Among those,
there can be inconsistencies that lead to ambiguity,
but we found that these only occur in instances where
multiple laws are referenced in one citation. This ref-
erence pattern is easily recognisable by two leading
article signs (”§§”) instead of one. However, in our
dataset, this pattern only occurs in 7.34% of all cita-
tions and a manual inspection of 100 of these patterns
showed only one instance of ambiguity.
Overall we managed to extract 1,279,105 case-to-
case citations (avg. 6.34 citations per court decision)
and 2,234,934 case-to-law citations (avg. 11.1 cita-
tions per court decision).
4 LEGAL CITATION NETWORK
The result of the data collection and citation extrac-
tion step is our legal citation network. In contrast to
previous research, we present a case-to-case citation
network that also connects cases to the laws that are
referenced within their decision text.
Figure 1: Neo4j graph representation of the legal citation
network.
Figure 1 shows the graph representation of the net-
Analysis of a German Legal Citation Network
149
work within a Neo4j database. The nodes of the graph
are ”Case”, ”Court” and ”Law”. Table 1 provides
an overview of the most important properties of each
node. Exactly one directed edge exists between two
Case nodes (n) and (m) if (n) references (m) in the
decision text at least once. For our analysis it is of
importance that a reference occurs, but not how of-
ten. Hence, multiple references from the same deci-
sion text to the same node are disregarded. Similarly,
an edge between a Case node (n) and Law node (l)
is created if (l) is cited at least once within the de-
cision text of (n). Each Case node is also connected
to the corresponding Court node that made the deci-
sion. Additionally, an edge between two Court nodes
(c) and (d) is created whenever an indirect citation be-
tween courts is identified. This occurs, when a Case
node (n), belonging to Court (c), is referencing a Case
node (m), belonging to Court (d).
Table 1: Node properties of the legal citation graph.
Node Property Example
Case DecisionText Der Antrag des Antrag-
stellers, § 1 Abs. 5
Corona VV HE 4 im
Wege der einstweiligen
Anordnung...
Case File Number IX ZR 70/20
Case Decision Date 25.03.2021
Law Article § 242
Law Statute BGB
Law Law Text Der Schuldner ist
verpflichtet, die Leistung
so zu bewirken...
Court Name Finanzgericht Hamburg
Court State Hamburg
Court Jurisdiction Finanzgerichtsbarkeit
As we can only add an edge between nodes that
exist in the database, extracted references to laws
or court decisions that are not yet included in the
database are ignored. The total number of nodes and
edges in the database are listed in table 2. This shows
that 59.9% of extracted references to laws could be
added as edges, while only 16.3% of references to
other court decisions are represented in the graph.
Those missed edges are mainly due to the lack of data
in the graph, as we made a particular effort to reduce
the chances of false-positive citation identification. A
copy of the database can be downloaded below
9
.
9
https://osf.io/8d2v4/
Table 2: Legal citation graph statistics.
Name Type Amount
Case Node 201,823
Law Node 50,814
Court Node 1,119
CASE REF Edge 208,585
LAW REF Edge 1,340,506
COURT REF Edge 7,612
5 RESULTS OF THE ANALYSIS
Network algorithms and properties are valuable tools
for exploring the behaviour of an interconnected sys-
tem. Even without being experts in the legal domain,
we can investigate these attributes to reveal and infer
information about the nature of German law.
5.1 Scale-free Network
Previous research has shown that the case citation net-
works of the U.S. Supreme Court (Smith, 2005), the
Austrian Supreme Court (Geist, 2009) and the Euro-
pean Court of Justice (Staffan Malmgren, 2011) ex-
hibit scale-free characteristics. The original hypothe-
sis for this fact stems from the similarity of case ci-
tation networks to the World Wide Web. Both are di-
rected and dynamically growing networks with large
hubs that tend to follow the principle of preferential
attachment (Barabasi, 2003). In the World Wide Web,
this principle has the effect that as the network grows,
websites with more links (”hubs”) are more likely to
receive new links than websites with fewer links.
Seeing the same behaviour in the German case
citation network would reveal that there is a very
small cluster of court decisions that hold a substan-
tial amount of legal influence. Similar to the works
above, we tested this hypothesis by inspecting the in-
degree distribution (number of incoming citations) of
our network. As evident by the histogram in figure 2,
our network displays the same typical power-law de-
gree distribution as other scale-free networks
10
. More
than 70% of court decisions are not cited at all and
92.6% of cases are cited less than five times, mean-
ing that only a small number of cases receive most
of the citations. Therefore, we can assume that the
German case citation network is following a scale-
free behaviour. The same conclusion can be made
when looking at the case-to-law in-degree histogram
in figure 3. Evidently, case and law references are not
10
As shown by (Lux et al., 2007), this type of diagram-
ming might not provide conclusive evidence but is sufficient
for this estimate and the scope of this work.
KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval
150
equally distributed but there are in fact ”hub”-like de-
cisions and laws that are more likely to be cited.
0 50 100 150 200 250 300
0
1
2
3
4
5
Case In-Degree Histogram
degree
# cases (log scale)
Figure 2: In-Degree distribution with log scale. The hor-
izontal axis shows the number of incoming citations (in-
degree value) and the vertical axis represents the corre-
sponding number of court cases. This type of distribution
suggests a scale-free network behaviour.
0 1 2 3 4 5
0
1
2
3
4
Law In-Degree Histogram
degree (log scale)
# laws (log scale)
Figure 3: In-Degree distribution of laws with log scale. This
scatter plot shows that most laws have not been cited at all,
while a small number of laws receive most of the citations.
Similar to figure 2, this type of distribution suggest a scale-
free network behaviour.
5.2 Centrality
One of the driving forces behind the development of
legal citation indexes and networks was the desire to
facilitate the search for important or influential court
decisions and precedents. Using citation counts as a
measure for excellence is a common practice in many
areas, including scientific research. Consequently, ef-
forts into analysing ranking scores based on different
citation metrics have been made (Geist, 2009; Staffan
Malmgren, 2011; Fowler et al., 2007; R
¨
onneburg,
2021). In-degree and PageRank scores have shown
to be strong indicators for identifying precedents or
otherwise influential cases.
5.2.1 PageRank
In figure 4 we identified the twenty most impor-
tant court decisions based on their overall (current)
PageRank rating. The stacked bars show the change
of their PageRank value over time (2005-2021). The
graph indicates that most decisions become highly in-
fluential after the third year of their appearance. One
exception seems to be the decision ”9 AZR 44/09”,
which denotes its highest PageRank value in the same
year of its existence.
1.73
1.94
3.66
3.15
6.65
5.44
5.12
7.98
5.96
7.64
12.83
18.79
11.03
11.15
12.11
27.34
4.71
28.87
47.7
1.77
2.98
15.29
2.47
21.16
16.62
6.53
4.67
5.87
33.04
15.27
9.97
22.63
48.04
8.26
20.66
11.06
48.78
6.67
7.51
7.86
5.27
10.57
11.57
12
22.85
15.39
28.46
4.97
37.4
7.92
6.11
12.68
12.19
13.24
50.61
24.76
30.68
12.61
21.9
40.4
8.23
13
15.48
18.72
23.1
16.78
15.76
15.69
35.23
16.59
23.33
21.73
8.57
50.73
22.97
13.46
17.11
24.48
24.76
15.44
21.98
1.74
42.36
16.51
23.98
17.24
35.14
25.18
17.81
17.02
16.67
26.27
21.5
23.15
20.02
15.76
22.51
15.82
43.6
2.15
50.7
17.84
17.11
16.32
19.52
16.04
20.02
18.26
16.44
7.42
51.12
16.66
35.04
43.82
26
23.12
21.17
22.86
23.96
28.28
23.96
18.41
23.42
20.03
21.33
20.19
19.01
16.89
23.05
16.49
15.34
43.98
16.03
34.93
16.42
28.34
26.08
50.89
17.09
11.78
20.28
23.96
23.05
28.53
26.34
17.04
16.49
17.36
15.98
16.49
15.37
44.03
21.6
34.84
19.94
16.37
19.38
23.6
19.29
50.88
16.37
26.34
16.49
23.05
34.84
16.49
19.38
19.29
19.94
15.37
15.98
44.03
17.04
21.6
17.36
23.6
50.88
28.53
20.28
23.96
16 E 550/09
2 AZR 541/09
16 B 332/07
1 A 21/12.A
IV ZR 76/11
9 AZR 44/09
C-394/12
1 ABR 19/10
1 A 1925/09
C-4/11
VGH B 35/12
10 C 5/09
A 11 S 1721/13
5 AZR 347/11
C-137/14
C-378/14
10 S 2182/04
B 12 KR 25/10 R
4 A 2103/15.A
IV ZR 73/13
0
100
200
300
400
500
2005
2010
2015
2020
year
Top 20 Court Decisions by PageRank
file number
PageRank
Figure 4: Stacked bar chart displaying the most influential
court decisions based on their PageRank value. Each bar
shows the development of the PageRank value over time.
B 14 AS 50/10 R
B 12 R 17/09 R
14 A 2708/10.A
C-209/12
IV ZR 76/11
C-378/14
4 A 2103/15.A
2 LB 91/17
C-163/17
VIII ZR 225/17
3 MR 10/20
B 4 AS 49/09 R
B 4 AS 99/10 R
1 B 234/12.A
C-394/12
VGH B 35/12
C-137/14
3 LB 17/16
A 11 S 2151/16
C-181/16
13 KN 510/18
2 MN 379/19
2 AZR 541/09
B 12 R 13/09 R
B 12 KR 25/10 R
C-4/11
1 A 21/12.A
4 B 1480/14
2 BvR 31/14
2 A 215/17
A 11 S 924/17
4 B 1333/18
13 B 539/20.NE
2010 2012 2014 2016 2018 2020
0
5
10
15
20
Fist
Second
Third
Cases with the highest PageRank by year
PageRank
Figure 5: Top three court decisions based on PageRank by
year. Older cases have a larger PageRank value as they have
more chances of being cited.
Similarly, figure 5 shows the top decisions based
on PageRank, but with respect to the year of the case.
In this particular diagram, we focused on the top three
decisions of the last ten years. This visualisation
shows that in some years the most important deci-
sions were made by the European Court of Justice.
For example, the case ”C-209/12” which is about im-
migration and freedom of settlement, has the highest
PageRank value of all decisions from 2013.
Analysis of a German Legal Citation Network
151
5.2.2 In-degree
Although (Staffan Malmgren, 2011) have shown that
rankings using PageRank can be more accurate than
using the in-degree value, it can still convey interest-
ing information. In figure 6 we have identified the
most cited court decisions of the last 40 years. It
is noticeable that some decisions maintain their im-
portance for consecutive years until a new ”breakout”
case receives more citations.
file number
C-163/17
2 LB 91/17
3 AZR 267/14
IV ZR 76/11
1 A 21/12.A
5 AZR 347/11
16 E 550/09
L 1 B 320/05 SF SK
13 B 1215/07
4 B 961/06
11 S 2779/04
6 B 2451/03
8 A 4782/99.A
8 A 1292/96.A
15 A 1047/99
9 A 5709/97
Figure 6: Court decisions with the highest amount of in-
coming citations per year. Some of the top cited cases seem
to stay important for a consecutive year, but do not receive
the same amount of attention after that.
The in-degree measure can also help to visualise
the citing behaviour between courts. Figure 7 shows
the most cited federal courts in Germany. The chart
shows that decisions form the federal social court
(”Bundessozialgericht”) are cited most often, while
the federal constitutional court (”Bundesverfassungs-
gericht”) received only 3.2% of all citations.
28.43
25.56
19.78
11.61
11.42
3.2
Bundessozialgericht
Bundesarbeitsgericht
Bundesgerichtshof
Bundesverwaltungsgericht
Bundesfinanzhof
Bundesverfassungsgericht
0
1000
2000
3000
4000
5000
6000
Most Cited Federal Courts
Figure 7: Among the federal courts, the social, labour and
justice courts receive the highest amount of citations. The
administrative and fiscal court receive each about 11% of all
citations, while the constitutional court is cited least often.
Another avenue of utilising the in-degree measure
is to find the most cited laws. In particular, it might
be interesting to visualise the most important laws of
each statute. As an example, figure 8 shows the most
cited laws of the German civil code (”BGB”).
3.79
3.73
2.83
2.74
2.57
2.13
1.86
1.72
1.71
1.67
1.55
1.54
1.48
1.1
1.03
1.03
0.99
0.95
0.91
0.87
§ 242
§ 288
§ 286
§ 280
§ 823
§ 133
§ 291
§ 157
§ 611
§ 812
§ 249
§ 307
§ 305
§ 254
§ 199
§ 195
§ 626
§ 134
§ 613
§ 1004
0
2000
4000
6000
8000
Most Cited Laws (top 20) in BGB
Article
In-degree
Figure 8: Most cited laws of the German civil code. Article
242 has been referenced most often receiving 3.79% of all
citations towards the civil code (BGB). Article 242 states:
”Performance in good faith: An obligor has a duty to per-
form according to the requirements of good faith, taking
customary practice into consideration.”.
5.3 Similarity
One important task of legal research includes the dis-
covery of related cases based on topic or field of law.
Previous research into legal document similarity (Ku-
mar et al., 2011; Wagh and Anand, 2017; Mandal
et al., 2017; Bhattacharya et al., 2020) has shown
that network-based similarity measures can perform,
but are also very perceptive to the sparsity of the net-
work. However, these findings are based on case-to-
case networks, the introduction of links to laws should
improve the accuracy of these measures, especially
for sparser case networks. However, the lack of la-
belled data is again aggravating the testing of this hy-
pothesis.
5.3.1 Node Similarity
To identify similar court decisions we compute the
Jaccard similarity score between all pairs of nodes
within the Neo4j graph. The algorithm considers a
pair of nodes as similar if they share the majority of
their neighbours. We compare the results of the algo-
rithm with a simplified text-based similarity calcula-
tion. For this, we determine the cosine similarity be-
tween the respective TF-IDF vectors of the decision
text pairs (without stop words). Table 3 shows some
examples of this comparison. As indicated by these
examples, the text-based and the network-based simi-
larity scores seem to be in agreement more often than
KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval
152
Table 3: Comparison between node similarity and TF-IDF based text-similarity.
Case1 Case2 Node Similarity TF-IDF Similarity
VI-3 Kart 18/09 (V) VI-3 Kart 17/09 (V) 1.00 1.00
VI-3 Kart 18/09 (V) VI-3 Kart 26/09 (V) 1.00 1.00
VI-3 Kart 18/09 (V) VI-3 Kart 27/09 (V) 1.00 0.99
VI-3 Kart 18/09 (V) VI-3 Kart 28/09 (V) 1.00 0.99
VI-3 Kart 18/09 (V) VI-3 Kart 29/09 (V) 1.00 0.99
W 7 M 19.30082 W 7 M 19.30083 1.00 0.96
W 7 M 19.30082 5 L 1635/14.TR 0.88 0.69
W 7 M 19.30082 17 L 1610/14.A 0.81 0.71
W 7 M 19.30082 7 L 1224/14.A 0.58 0.55
W 7 M 19.30082 3 E 187/17 0.50 0.53
L 8 R 208/05 L 8 R 361/06 1.00 0.98
L 8 R 208/05 L 8 R 44/06 0.86 0.97
L 8 R 208/05 L 8 R 47/06 0.86 0.13
L 8 R 208/05 L 8 R 62/07 0.80 0.97
L 8 R 208/05 L 3 R 98/05 0.35 0.97
not. In order to quantify this agreement, we calculated
the Pearson correlation between the similarity mea-
sures of 1,000 semi-randomly chosen court decisions
and their five most similar counterparts. To avoid dis-
connected or unique cases we randomly selected these
samples from a list of cases that contain at least ten ci-
tations and one similar case (i.e. at least one pair of
nodes with a 0.5 Jaccard similarity score). The re-
sult of 0.64 implies a positive correlation, which sug-
gests that network-based similarity can provide an ad-
ditional avenue for case-similarity search methods.
6 CONCLUSION & FUTURE
WORK
In this paper, we introduced the development of a
German legal citation network containing over 50,000
laws and 200,000 decisions from more than 1000
courts from all levels of appeal by using the Open Le-
gal Data dataset (Ostendorff et al., 2020). In North
America the analysis of case-to-case citation net-
works for legal information retrieval purposes has a
much long research history, compared to most Euro-
pean countries. One factor for this discrepancy orig-
inates from the lack of openly available legal data in
Europe and Germany in particular. In this context,
we highlighted the difference between the extensive-
ness of North American open legal data platforms and
their German counterparts. In extension, we also un-
derlined that the lack of annotated data aggravates the
development of citation extraction approaches. How-
ever, we found that a rule-based approach using regu-
lar expressions can work sufficiently well. We stored
the resulting network in a Neo4j database to enable
efficient querying and the use of graph algorithms.
Using in-degree histograms, we show that the Ger-
man case-to-case and case-to-law citation networks
exhibit scale-free behaviour. We use the PageRank al-
gorithm and in-degree scores to reveal the most influ-
ential court decisions, laws and courts (we have taken
the federal courts as an example). Lastly, we indicate
that node similarity measures have a positive correla-
tion to text-based similarity scores.
Future research in this domain could explore the
impact of citation behaviour on the outcome of a court
decision. For this, it would be necessary to enrich
the database with a label for the decision’s outcome.
Fortunately, German court decisions are well struc-
tured and should support automatic classification ef-
forts. Furthermore, community detection algorithms
could be used to help identify laws or cases that are
often cited together. Lastly, if there was more access
to annotated citations from court decisions and laws,
improvements could be made to the citation extrac-
tion task and more links could be added to the net-
work. This could also include expanding the network
to include links between laws.
ACKNOWLEDGEMENTS
The project on which this report
is based was funded by the Ger-
man Federal Ministry of Educa-
tion and Research (BMBF) under
the funding code 01—S20049.
The author is responsible for the
content of this publication.
Analysis of a German Legal Citation Network
153
REFERENCES
Barabasi, A.-L. (2003). The New Science of Networks. J.
Artificial Societies and Social Simulation, 6.
Bhattacharya, P., Ghosh, K., Pal, A., and Ghosh, S. (2020).
Methods for Computing Legal Document Similarity:
A Comparative Study. In LDA 2019 Workshop in the
JURIX 2019 Conference.
Corinna Coupette (2019). Juristische Netzwerkforschung.
Mohr Siebeck.
Cross, F., Spriggs, J., Johnson, T., and Wahlbeck, P. (2010).
Citations in the U.S. Supreme Court: An Empirical
Study of Their Use and Significance. University of
Illinois law review, pages 489–575.
Fowler, J., Johnson, T., Spriggs, J., Jeon, S., and Wahlbeck,
P. (2007). Network Analysis and the Law: Measur-
ing the Legal Importance of Precedents at the U.S.
Supreme Court. Political Analysis, 15.
Geist, A. (2009). Using Citation Analysis Techniques for
Computer-Assisted Legal Research in Continental Ju-
risdictions. SSRN Electronic Journal.
Gelter, M. and Siems, M. (2012). Networks, Dialogue or
One-Way Traffic? An Empirical Analysis of Cross-
Citations Between Ten of Europe’s Highest Courts.
Utrecht Law Review, 8(2).
Katz, D. M., Coupette, C., Beckedorf, J., and Hartung, D.
(2020). Complex societies and the growth of the law.
Scientific Reports, 10(1).
Kumar, S., Reddy, P. K., Reddy, V. B., and Singh, A. (2011).
Similarity Analysis of Legal Judgments. In Proceed-
ings of the Fourth Annual ACM Bangalore Confer-
ence, COMPUTE ’11, New York, NY, USA. Asso-
ciation for Computing Machinery.
Leitner, E., Rehm, G., and Schneider, J. M. (2020). A
Dataset of German Legal Documents for Named En-
tity Recognition. CoRR, abs/2003.13016.
Lux, M., Granitzer, M., and Kern, R. (2007). Aspects of
Broad Folksonomies. In 18th International Confer-
ence on Database and Expert Systems Applications
(DEXA 2007). IEEE.
Mandal, A., Chaki, R., Saha, S., Ghosh, K., Pal, A., and
Ghosh, S. (2017). Measuring Similarity among Le-
gal Court Case Documents. In Proceedings of the
10th Annual ACM India Compute Conference on ZZZ
- Compute ’17, New York, New York, USA. ACM
Press.
Milz, T. and Seifert, C. (2018). Analysing Author
Self-citations in Computer Science Publications. In
Database and Expert Systems Applications, pages
289–300, Cham. Springer International Publishing.
Neale, T. (2013). Citation Analysis of Canadian Case Law.
Journal of Open Access to Law, vol. 1, no. 1, pp. 1–51.
Ostendorff, M., Blume, T., and Ostendorff, S. (2020). To-
wards an Open Platform for Legal Information. Pro-
ceedings of the ACM/IEEE Joint Conference on Digi-
tal Libraries in 2020.
R
¨
onneburg, L. (2021). Analyse eines juristischen Entschei-
dungscorpus mit Methoden der Netzwerkforschung
und Sprachtechnologie. University of Hamburg,
Bachelor Thesis.
Smith, T. A. (2005). The Web of Law. SSRN Electronic
Journal.
Staffan Malmgren (2011). Towards a theory of jurispruden-
tial relevance ranking Using link analysis on EU case
law. Stockholm University, Master Thesis.
Urchs, S., Mitrovic, J., and Granitzer, M. (2020). Towards
Classifying Parts of German Legal Writing Styles in
German Legal Judgments. In 2020 10th International
Conference on Advanced Computer Information Tech-
nologies (ACIT). IEEE.
Urchs, S., Mitrovi
´
c, J., and Granitzer, M. (2021). De-
sign and Implementation of German Legal Deci-
sion Corpora. In Proceedings of the 13th Interna-
tional Conference on Agents and Artificial Intelli-
gence. SCITEPRESS - Science and Technology Pub-
lications.
Wagh, R. and Anand, D. (2017). Application of citation
network analysis for improved similarity index esti-
mation of legal case documents : A study. In 2017
IEEE International Conference on Current Trends in
Advanced Computing (ICCTAC), pages 1–5.
Widdison, R. (2002). New Perspectives in Legal Informa-
tion Retrieval. International Journal of Law and In-
formation Technology, 10(1):41–70.
KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval
154